Introduction

Research background and motivations

With the rapid development of online education, personalized learning platform has gradually become the key technology to improve learning effect and students’ participation. In traditional music learning, the limitations of teachers and textbooks make students lack flexibility in learning path and content selection, and it is difficult to meet each student’s interest and learning progress1,2,3. Driven by Artificial Intelligence (AI) and deep learning technology, the intelligent online music learning platform provides students with a more personalized learning experience, especially in the course recommendation system, which can accurately recommend students according to their interests and historical learning behaviors, thus promoting their learning motivation and effectiveness4,5,6.

However, most of the existing music learning platforms adopt static recommendation algorithms, which are difficult to capture the dynamic changes of students’ interests in real time, and fail to effectively integrate students’ historical behavior and current learning status7,8. With the continuous evolution of students’ interests and needs, the traditional recommendation methods based on user portraits or collaborative filtering cannot adapt to this change. It often leads to the mismatch between the recommended course content and students’ real interests, which affects the learning experience and learning effect9,10. Therefore, how to design a dynamic and intelligent recommendation system in the music learning platform, accurately capture the evolution of students’ interests and provide personalized course recommendation according to its changes has become the key issue of current research.

Research objectives

The main objective of this study is to design a student-centered AI online music learning platform, the core of which is the course recommendation model based on the evolution of students’ interests. Specifically, the objective is achieved in the following aspects: (1) Build a student-centered AI online music learning platform, capture the dynamic changes of students’ interests through deep learning, and recommend relevant courses for them. (2) Design an efficient input module to embed and process students’ multidimensional data (including user portraits, historical behaviors and course features, etc.) to provide effective input for subsequent models. (3) Combine Gated Recurrent Unit (GRU) and attention mechanism, improve the learning ability of the model to students’ historical behavior, and optimize the recommendation results by modeling the co-occurrence relationship among courses with project crossing module. This study aims to provide accurate and efficient course recommendation scheme for online music learning platform and improve students’ learning experience.

Literature review

Application of deep learning in online music learning platform

With the rapid development of AI technology, deep learning has become an important research tool in many fields, and also plays a key role in the research of online music learning platform. In recent years, researchers have explored AI-driven methods to enhance personalized learning experience and optimize the accuracy of recommendation systems. Jin and Zhang (2024) proposed a music online education resource recommendation system combining blockchain and deep learning. The system integrated various algorithms to improve the personalized level and accuracy of recommendation. The experimental results showed that the recommendation accuracy of the improved collaborative filtering algorithm was as high as 95%. The expert evaluation showed that the system was superior to the traditional method in recommendation quality, which had high application value and could be widely used in online music education platform11. The study enhances the transparency and security of the recommendation system through blockchain technology. It also optimizes personalized recommendations using deep learning, improving both accuracy and interpretability. However, this method has high computational costs, making it difficult to adapt to real-time updates of large-scale user data. Additionally, the integration of blockchain technology requires extra storage and computational resources, which may limit its practical application in large-scale online education platforms. Jiang (2023) used cloud computing and fuzzy C-means clustering technology to optimize online music learning, classified students individually through data analysis, and matched them with suitable learning resources, and built a music learning platform that supported 400 people to be online at the same time. The experimental results showed that the system was excellent in scalability, availability and effectiveness, and the normal operation rate was as high as 99%, showing good stability and practical application value12. The study’s strength lies in using cloud computing to enhance the system’s computational power. This allows personalized recommendations to meet the needs of large-scale users. Meanwhile, fuzzy C-means clustering effectively handles the uncertainty of user interests. However, the method relies on high-quality initial data classification. If the initial clustering is inaccurate, it may affect the recommendation results. Additionally, the system mainly focuses on personalized classification. It does not deeply explore the dynamic evolution of user interests. Therefore, the adaptability of recommendations may be insufficient in long-term learning processes. Liu and Shao (2024) developed an online piano course consisting of three parts: theoretical study, technical training and skill consolidation. The survey results showed that among various online piano learning schemes, Xiaoye AI piano teacher’s learning effect was the most significant, and 29% of the students surveyed thought that his teaching was the most effective, especially in polyphony training. The research further showed that the introduction of digital courses had effectively improved students’ playing skills and polyphonic ability, and fully verified the feasibility and effectiveness of online piano learning13. The study highlights the potential of artificial intelligence in music skill training. Specifically, in personalized teaching, its deep learning-based feedback system can adjust teaching content in real time to suit different students’ learning progress and needs. However, the system mainly focuses on piano teaching, limiting its scope of application. It has not yet been extended to other music learning fields. Additionally, the study lacks long-term progress tracking analysis of learners. In the future, combining more behavioral data could optimize the personalized teaching model, further enhancing learning outcomes.

Overall, these studies discuss the application of AI technology in online music learning from different angles, covering intelligent recommendation, personalized learning path construction, digital curriculum development and many other aspects. They provide important theoretical support and practical basis for building a more efficient and accurate online music learning platform. However, the existing research still has some limitations in interest dynamic modeling, computational overhead optimization and the adaptability of recommendation system.

Application of personalized recommendation system in online education

Personalized recommendation system is an important application of deep learning in the field of education, especially in the recommendation of learning resources and course selection. Deep learning model can significantly improve the accuracy of the system and user satisfaction. Zhang et al. (2021) discussed the main recommendation technologies used in e-learning, including content-based, collaborative filtering and knowledge-based methods, and emphasized the application of these technologies to meet specific needs in e-learning14. The study shows that content-based recommendations can match course text information, while collaborative filtering methods recommend based on users’ historical behaviors. However, content-based methods tend to fall into the “information silo” problem, failing to effectively recommend new courses users have not encountered. Collaborative filtering methods have limitations in cold-start scenarios, making it difficult to provide effective recommendations for new users. Additionally, knowledge-based methods can combine educational ___domain knowledge graphs for precise recommendations, but they rely on high-quality knowledge base construction, limiting their application scope. Therefore, although the study systematically summarizes different recommendation techniques, it does not address how to optimize recommendation effectiveness in dynamic environments. Bhaskaran and Marappan (2023) designed a new hybrid personalized recommendation system based on inductive support vector machine (SVM) to improve the recommendation quality of machine learning public datasets. By analyzing learners’ habits, the proposed model performed well in recommendation tasks in many fields15. However, the study primarily experiments with static datasets and does not consider the dynamic evolution of student interests in online education environments. Additionally, the ISVM model relies on predefined rules, limiting its adaptability. It struggles to flexibly adjust to meet the personalized needs of different users. Therefore, although the study provides a new method for personalized recommendations, its applicability in large-scale dynamic learning environments still requires further validation. Javed et al. (2021) used content-based recommendation, collaborative filtering and mixed recommendation system, and combined with Web ontology language and resource description framework to provide personalized news and e-learning content recommendation16. The system effectively improves user learning efficiency in e-learning environments and enhances recommendation interpretability through semantic technologies. However, this method has high requirements for data quality, relying on complete and structured knowledge bases. In reality, educational data is often unstructured, leading to certain limitations when the system processes complex educational resources. Additionally, the study primarily focuses on semantic-based recommendation methods and does not explore the application of deep learning in dynamic modeling of user interests. Therefore, there is still significant room for optimization in complex user behavior modeling and high-order feature interactions.

Despite significant progress in personalized recommendation systems in the field of online education, most existing studies still face the following issues: (1) Many models rely on static data and fail to fully consider the dynamic evolution of students’ learning interests; (2) Traditional collaborative filtering and content-based recommendation methods struggle to effectively capture long-term trends in student interests; (3) Existing recommendation systems still have considerable room for improvement in complex user behavior modeling and high-order feature interactions. To address these issues, this study proposes a deep learning-based course recommendation model. The model combines attention mechanisms and GRU, and introduces an item interaction module to better capture the dynamic evolution of student interests, improving recommendation accuracy and adaptability.

The latest progress of deep learning in related fields

In recent years, the application of deep learning in the field of recommendation system has been expanding, and remarkable achievements have been made in many related fields. Although some studies do not directly involve online education or curriculum recommendation, their theoretical methods and technical realization have important reference value for this study. For example, Zhao et al. (2024) proposed a key flow identification algorithm based on deep reinforcement learning, and combined with linear programming to optimize the traffic scheduling strategy of SRv6 network. The experimental results showed that this method effectively improved the load balancing ability, significantly reduced the end-to-end transmission delay, and enhanced the network’s ability to adapt to dynamic environmental changes17. Deep reinforcement learning can achieve efficient decision optimization in dynamically changing environments. This has certain implications for personalized recommendation systems in online education platforms. For example, students’ learning interests and behavior patterns change dynamically. Traditional recommendation systems struggle to adapt in real time. The adaptive optimization capabilities of deep reinforcement learning could be used to improve personalized recommendation systems. This would enable continuous optimization based on changes in student behavior, enhancing recommendation accuracy and user experience. Zhang et al. (2024) discussed the vulnerability types, fuzzy testing methods and the vulnerability sources of PyTorch and MLIR frameworks around the security issues of deep learning frameworks, and looked forward to future research directions, such as automatic error repair and LLM-enabled fuzzy testing18. The study highlights the security challenges of deep learning models in practical applications. In data-driven systems, they may be affected by adversarial attacks or data pollution. For online education recommendation systems, data security and model robustness are equally critical. If the recommendation system is compromised by data pollution or malicious manipulation, it may lead to unreasonable course recommendations, negatively impacting students’ learning experiences. Therefore, the study suggests that when building models, it is necessary to consider their security and stability. Methods such as adversarial training, anomaly detection, or enhancing model interpretability can improve the reliability and controllability of recommendation systems.

In summary, although the above studies do not directly address online music learning recommendation systems, their contributions in areas such as deep reinforcement learning for dynamic environment optimization and deep learning model security provide important insights for this research. The dynamic optimization methods of deep reinforcement learning can be used to improve the adaptability of personalized recommendation systems. Meanwhile, research on deep learning model security helps enhance the robustness and reliability of recommendation systems. These ideas offer valuable references for further optimizing recommendation systems in online music learning platforms in the future.

Research innovation and contribution of this study

Although existing research has made progress in personalized recommendation systems and online education, challenges remain in modeling the dynamic evolution of student interests and improving the adaptability of recommendation systems. For example, most current studies rely on static features and fail to fully utilize students’ historical learning behaviors for dynamic interest modeling. Additionally, traditional recommendation methods have limitations in handling high-order feature interactions, making it difficult to capture complex relationships between courses. To address these issues, this study proposes a Course Recommendation Model for Student Learning Interest Evolution (CRM-SLIE). The main innovations of this model include: (1) combining attention mechanisms and GRU to enhance the modeling of dynamic changes in student learning interests; (2) introducing an item interaction module to improve the system’s ability to capture high-order feature interactions; and (3) optimizing the model structure by experimentally validating the impact of different embedding dimensions and behavior sequence lengths on system performance, ensuring adaptability to various learning scenarios.

In summary, building on existing research, this study optimizes the core architecture of personalized recommendation systems by introducing deep learning methods. It also conducts an in-depth exploration of application scenarios in online music learning platforms. The goal is to improve the accuracy, adaptability, and stability of recommendation systems, providing strong support for building more intelligent online music learning platforms.

Research model

Overall architecture design of online music learning platform

The AI online music learning platform designed in this study aims to provide students with personalized learning experience, especially in the selection and recommendation of music courses. The platform architecture adopts modular design, which mainly includes course recommendation module, student data management module, learning path management module and real-time feedback module, as shown in Fig. 1.

Fig. 1
figure 1

Overall architecture diagram of AI online music learning platform.

Full size image

In Fig. 1, the student data management module collects students’ basic information, historical learning records, interest preferences and real-time learning behaviors to form a comprehensive student portrait. Through the evolution module of students’ interests, the platform can track the changes of students’ interests and learning progress to adjust the recommendation strategy in real time and ensure the timeliness and pertinence of the recommended content. Secondly, the learning path management module intelligently plans the learning path according to students’ ability development and course difficulty, and gradually guides students to complete their learning goals. The core part of the platform-course recommendation module, combined with deep learning technology, can generate personalized course recommendation according to students’ interests and historical behaviors. By learning students’ interactive behaviors, such as audition, comment and feedback, the module continuously optimizes the recommendation algorithm and realizes accurate recommendation. The platform also includes a real-time feedback module, which generates learning reports and feedback in real time through the interactive data between students and the platform, helps students understand their own learning strengths and weaknesses, and provides targeted learning suggestions when appropriate. Through the collaborative work of the above modules, a flexible and efficient AI music learning system is constructed, which can realize personalized recommendation and accurate guidance and provide students with the best learning resources and experience.

Design of learning course recommendation model based on students’ interest and deep learning

In the student-centered AI online music learning platform, the course recommendation module is the core part to realize personalized learning experience. In this study, a Course Recommendation Model for Student Learning Interest Evolution (CRM-SLIE) is proposed. The overall architecture includes five main modules: input module, GRU module, interest evolution module, project crossing module and output module, as shown in Fig. 2.

Fig. 2
figure 2

Schematic diagram of the overall structure of CRM-SLIE model.

Full size image

In the CRM-SLIE model shown in Fig. 2, each module collaborates through precise mechanisms to achieve personalized course recommendations. First, the input module processes students’ historical behaviors and basic information to generate a feature vector embedded with positional encoding. This vector serves as the input for subsequent modules, forming the foundational data of the model. Next, the GRU module processes the input feature vector to extract potential interest states based on students’ historical behavior sequences. Through its recursive mechanism, this module captures temporal changes, allowing students’ interests to evolve over time. Here, the input feature vector uses gating mechanisms to determine the influence of historical states on the current interest state. The output of the GRU module represents the student’s current interest state, providing a dynamic representation of learning interests for subsequent modules. The student interest evolution module builds on the interest state generated by the GRU module and introduces an attention mechanism to dynamically adjust the interest state. By calculating the similarity between the current interest state and candidate courses, the model identifies the most relevant interest features for the current learning task and assigns different weights to historical behaviors. The core role of the attention mechanism is to help the model focus on key learning interests while ignoring irrelevant historical information, effectively addressing the interest drift phenomenon. The item interaction module captures co-occurrence relationships between historical learning courses and candidate courses through feature interaction modeling. Its core function is to extract second-order interaction features between courses using inner product and Hadamard product operations, thereby enhancing recommendation accuracy. The interacted feature information is passed to the output module, further enriching the relationship between student interest states and course features. Finally, the output module aggregates the outputs of all modules. By concatenating the interest state, course features, and user profile features, the model inputs this information into a two-layer fully connected network. The final output is generated through the PReLU activation function and softmax, producing personalized course recommendation probabilities. In this way, the model can accurately recommend the most suitable learning resources based on students’ evolving interests and historical behaviors. In summary, the interaction of each module, through layered feature processing and optimization, enables the CRM-SLIE model to better simulate the dynamic evolution of students’ learning interests, ultimately providing accurate and personalized course recommendations. The principles of each module are as follows:

Input module

In CRM-SLIE model, the task of input module is to transform students’ input data into a form suitable for model learning. Input data includes discrete features (such as students’ gender and course category) and continuous features (such as students’ age and study time)19,20. These features are processed by embedding layer, which transforms discrete features into low-dimensional embedding vectors, reduces data complexity and maintains the correlation among features. Let the embedding vector of a discrete feature \(\:{s}_{t}\) be \(\:{e}_{t}\), and the calculation equation is:

$$\:{e}_{t}={Q}_{t}{s}_{t}$$
(1)

\(\:{e}_{t}\in\:{\mathbb{R}}^{nd}\) represents the embedding vector of the t-th feature. \(\:{s}_{t}\in\:{\mathbb{R}}^{ni}\) is the original input data. \(\:{Q}_{t}\in\:{\mathbb{R}}^{nd\times\:ni}\) is the learned embedding matrix.

After standardization, continuous features are spliced with embedded vectors of discrete features to form a comprehensive feature vector. In order to keep the time sequence information, historical behavior data also introduces Positional Encoding (PE), and uses sine and cosine functions to calculate the position encoding:

$$\:PE\left(p,2i\right)=\text{s}\text{i}\text{n}\left(\frac{p}{{10000}^{\frac{2i}{d}}}\right)$$
(2)
$$\:PE\left(p,2i+1\right)=\text{c}\text{o}\text{s}\left(\frac{p}{{10000}^{\frac{2i}{d}}}\right)$$
(3)

\(\:p\) is the project ___location. d is the ___location coding dimension, and i is the dimension index of the vector.

All features are spliced to obtain an input vector \(\:X\), which is expressed as:

$$\:X=\left[{e}_{1},{e}_{2},\cdots\:,{e}_{t}\right]\oplus\:\left[{pe}_{1},{pe}_{2},\cdots\:,{pe}_{t}\right]$$
(4)

\(\:{e}_{t}\) is the feature embedding vector. \(\:{pe}_{t}\) is the position coding vector, and \(\:\oplus\:\) represents the splicing operation of vectors. As the representation of each behavior item, the input vector \(\:X\) is introduced into the subsequent deep learning model to learn the interest state and recommend courses.

GRU module

After processing the input data, the model enters the GRU module, the core task of which is to extract the potential interest state according to the students’ historical learning records. GRU is a variant of Recursive Neural Network (RNN), which can effectively alleviate the gradient disappearance problem of traditional RNN in long sequence learning, and has a lower risk of over-fitting because of its less parameter requirements21,22. In addition, GRU has fewer parameters and simpler structure compared with long short-term memory (LSTM) networks, and can achieve better performance at lower computational cost. Therefore, GRU is especially suitable for dealing with a large number of historical learning records with time series characteristics, such as the student behavior sequence in this study. Through GRU module, the model can capture the time series changes of students’ interest in learning, effectively extract the potential interest state, provide accurate basis for personalized course recommendation, and then optimize students’ learning path and experience. In the course recommendation task, GRU extracts the learning interest of each behavior item by learning the time series characteristics of students’ historical behaviors. GRU is calculated by the following equations:

$$\:{c}_{t}=\sigma\:({W}_{c}{x}_{t}+{U}_{c}{h}_{t-1}+{b}_{c})$$
(5)
$$\:{r}_{t}=\sigma\:({W}_{r}{x}_{t}+{U}_{r}{h}_{t-1}+{b}_{r})$$
(6)
$$\:{\stackrel{\sim}{h}}_{t}=tanh({W}_{h}{x}_{t}+{c}_{t}\circ\:{U}_{h}{h}_{t-1}+{b}_{h})$$
(7)
$$\:{h}_{t}=\left(1-{r}_{t}\right)\circ\:{h}_{t-1}+{r}_{t}\circ\:{\stackrel{\sim}{h}}_{t}$$
(8)
$$\:\sigma\:\left({x}_{1},{x}_{2}\right)=\frac{1}{1+\text{exp}[-({x}_{1}{x}_{2}\left)\right]}$$
(9)

\(\:{c}_{t}\) is the reset gate. \(\:{r}_{t}\) is the update gate. \(\:{\stackrel{\sim}{h}}_{t}\) is the candidate hidden state. \(\:{h}_{t}\) is the final hidden state at the current moment. \(\:\sigma\:\) is sigmoid activation function. \(\:\circ\:\) represents Hadamard product. W and U represent weight matrix, and b is bias term.

Through GRU module, CRM-SLIE model can extract the temporal interest characteristics of students’ historical learning behavior, and generate accurate interest status as the basis for subsequent personalized course recommendation. These learned interest states can help the model to recommend courses for students more accurately and optimize their learning paths and experiences.

Interest evolution module

In the student-centered online music learning platform, students’ interest in learning is dynamically evolving, and with the passage of time, students’ interest will change23. Although GRU can effectively capture the time series characteristics, it does not fully consider the relationship between students’ current interests and candidate courses. In order to better capture the changes of students’ interests, an Attention-based GRU (AGRU) model combining attention mechanism and GRU is proposed. AGRU can update students’ interest status and weight them according to the correlation between current interest and candidate courses, so that the model can dynamically adjust the interest evolution path and strengthen the interest information related to the current learning task. The introduction of attention mechanism enables the model to focus on the key parts of students’ interest changes when dealing with historical behavior sequences, ignoring the irrelevant or less influential parts, and improves the adaptability and accuracy of the model to the evolution of students’ interest. Specifically, let the sequence of students’ interest states be \(\:I=\{{i}_{1},{i}_{2},\cdots\:,{i}_{t}\}\), where \(\:{i}_{t}\) represents the students’ interest states at the t moment. \(\:{z}_{j}\) is the j-th candidate course. By introducing the attention mechanism, the model calculates the similarity between the current interest state and each candidate course, and obtains an attention score \(\:{a}_{t}\):

$$\:{a}_{t}=\frac{\text{exp}\left({i}_{t}W{z}_{j}\right)}{\sum\:_{k=1}^{T}\text{exp}\left({i}_{k}W{z}_{j}\right)}$$
(10)

\(\:{i}_{t}\in\:{\mathbb{R}}^{{n}_{I}}\) is the interest state vector at the t moment. \(\:{z}_{j}\in\:{\mathbb{R}}^{{n}_{z}}\) is the embedding vector of the candidate courses. \(\:W\in\:{\mathbb{R}}^{{n}_{I}\times\:{n}_{z}}\) is the weight matrix. \(\:{n}_{I}\) and \(\:{n}_{z}\) are the dimensions of interest status and candidate courses, respectively.

Different from the traditional attention mechanism, this study adopts the attention score without normalization, and retains the importance of each historical record, instead of mapping it to a standardized value between 0 and 1. In this way, the model can better retain the part of strong interest in historical records and avoid the important information that may be lost when dealing with long historical series. After calculating the attention scores, the AGRU structure embeds these scores into the update gate of GRU to adjust the students’ interest state more finely:

$$\:{u}_{t}={a}_{t}\times\:{r}_{t}$$
(11)

\(\:{r}_{t}\) is the update gate in GRU. \(\:{u}_{t}\) is the weighted update gate. By combining this weighted update gate with GRU’s standard update rules, a new interest state is generated:

$$\:{h}_{t}^{{\prime\:}}=\left(1-{u}_{t}\right)\circ\:{h}_{t-1}+{u}_{t}\circ\:{\stackrel{\sim}{h}}_{t}$$
(12)

\(\:{h}_{t-1}\) is the interest state at the last moment. \(\:{\stackrel{\sim}{h}}_{t}\) is the hidden state of the current candidate courses. \(\:{h}_{t}^{{\prime\:}}\) is the interest state after the AGRU update. Through the AGRU structure, the model can dynamically select the historical interest state related to the current learning task, reduce the interference unrelated to the current learning goal, effectively deal with the phenomenon of “interest drift”, and improve the accuracy and personalization level of the recommendation system.

Project crossing module

In the online music learning platform, a key challenge of course recommendation system is to model the relationship among course resources, especially the co-occurrence relationship between courses, that is, in the historical learning behavior, two courses often appear in the learning sequence at the same time24. In order to identify students’ potential interests and provide accurate recommendations, a project crossing module is proposed to capture the second-order characteristic interaction between courses and candidate courses. This module adopts bilinear crossing method, combining inner product and Hadamard product to extract complex interactive information among courses. Specifically, let the embedding vector of the i-th history learning course be \(\:{l}_{t}\), which is crossed with the embedding vector \(\:{z}_{j}\) of the candidate course:

$$\:{v}_{t,j}={l}_{t}\cdot\:{W}_{tj}\circ\:{z}_{j}$$
(13)

\(\:{W}_{tj}\in\:{\mathbb{R}}^{{n}_{I}\times\:{n}_{z}}\) is a parameter matrix used to capture the intersection of second-order features. \(\:\cdot\:\) represents the inner product operation. A new candidate course feature vector \(\:{v}_{j}\) is generated by averaging the results of feature crossover:

$$\:{v}_{j}=avg\left({v}_{t,j}\right)$$
(14)

This module can effectively model the co-occurrence relationship among course resources, especially in sparse data environment, and can tap the potential value of low-frequency features to avoid the problem of information loss, thus improving the accuracy and effect of the recommendation system and providing students with more accurate recommendations of learning resources.

Output module

In the CRM-SLIE model, the output module undertakes the key task of synthesizing the feature information generated by each module and generating the final recommendation result. In the previous steps, the interest evolution module generated the final learning interest vector \(\:{h}_{t}^{{\prime\:}}\). The project crossing module generates the feature vector \(\:{v}_{j}\) of the candidate courses. These features are input to the output module together with the user portrait feature \(\:u\) in series to further capture the nonlinear combination relationship of course recommendation. Specifically, the feature vector \(\:M\) after series connection is as follows:

$$\:M=Concatenate[{h}_{t}^{{\prime\:}},{v}_{j},u]$$
(15)

In this operation, three input vectors are spliced into a dense feature vector \(\:M\), which is input into the softmax function, and the probability of students browsing candidate courses is obtained. In order to enhance the nonlinear expression ability of the model, the activation function uses the Parametric Rectified Linear Unit (PReLU) function, and its specific form is:

$$\:PReLU\left({x}_{i}\right)=\left\{\begin{array}{c}{x}_{i},if\:{x}_{i}>0\\\:\gamma\:{x}_{i},if\:{x}_{i}\le\:0\end{array}\right.$$
(16)

\(\:\gamma\:\) is an adaptive parameter. This activation function can help the model better capture the negative activation information and improve the learning ability of the model. After activation by PReLU, the feature vector \(\:M\) enters the two-layer fully connected network for further processing:

$$\:{L}_{1}=PReLU({W}_{1}\cdot\:M+{b}_{1})$$
(17)
$$\:{L}_{2}=PReLU({W}_{2}\cdot\:{L}_{1}+{b}_{2})$$
(18)

\(\:{L}_{1}\) and \(\:{L}_{2}\) are the outputs of the fully connected layer. Finally, after the second level of full connection, the output vector \(\:{L}_{2}\) is transformed into the probability distribution of course recommendation by the softmax function:

$$\:y=softmax\left({L}_{2}\right)$$
(19)

The output represents the probability that the user selects each candidate course. The recommendation system will eventually select the course item set with the highest probability for personalized recommendation.

In order to optimize the learning process of the output module, the model adopts Binary Cross-Entropy as the loss function, and the specific form is25:

$$\:Loss=-\frac{1}{N}\sum\:_{i=1}^{N}[{y}_{i}\text{log}{\widehat{y}}_{i}+\left(1-{y}_{i}\right)\text{log}\left(1-{\widehat{y}}_{i}\right)]$$
(20)

\(\:{y}_{i}\) is the actual label of training sample i. \(\:{\widehat{y}}_{i}\) is the prediction probability of this sample by the model, and \(\:N\) represents the number of training samples. Finally, the output module combines the learning results of multiple modules to provide students with accurate music learning resources recommendation and promote the realization of personalized learning.

To sum up, the steps to summarize the CRM-SLIE model are as follows in Table 1:

Table 1 Implementation steps of CRM-SLIE model algorithm.
Full size table

Experimental design and performance evaluation

Datasets collection

This study uses the Canvas Network dataset published by Harvard Dataverse data website in 2016, which is widely used in the research of educational data mining and recommendation system26. This dataset contains many kinds of information, including the basic attributes of online courses, user portraits and user interaction behaviors, such as browsing, clicking, collecting and purchasing, which provides reliable data support for building an efficient personalized recommendation system. In order to ensure the correlation between the data and the research goal, the courses related to music education in the dataset are screened, and finally a sub-dataset containing 783 music courses is constructed. These courses cover many music learning fields, including but not limited to music theory, piano playing, guitar skills, vocal music training, composition and arrangement, etc., which provide a good data basis for studying the applicability of personalized recommendation system in different music learning scenarios. In the aspect of sample selection, in order to ensure the quality of data and the reliability of experiments, strict screening criteria are formulated. Firstly, in course selection, only courses with high user interaction are reserved, that is, courses with at least 50 users browsing or clicking records to ensure the validity of the data and ensure that the recommendation task can fully learn the correlation characteristics between courses. Secondly, in terms of user selection, in order to improve the generalization ability of the recommendation model, only active users who have participated in at least five courses and have a complete record of learning behavior are selected. This standard helps to eliminate those low-active users who have only a small amount of interactive data, thus reducing the impact of data noise on model training. In addition, in order to reduce the influence of data sparsity on the experimental results and ensure that the recommendation system can learn based on the latest interest trends, this study adopts the time interval screening standard, and only retains the interactive data in the past 12 months to ensure that the model can better adapt to the dynamically changing learning interests of users.

In the aspect of feature construction, 9 core features are selected in the experiment, including five attributes related to the course, such as course category, course difficulty, course duration, instructor, course score, age, study preference and study duration, and the user’s historical behavior sequence (including browsing, clicking and collecting). These features are pre-processed to improve the learning effect of the model. The main label in the experiment is whether the user browses a specific course, so that the recommendation model can be supervised and trained. For classification features, barrel coding is used to transform string features into discrete values for model processing. In addition, considering that the problem of time crossing may lead to the false high evaluation results of the model, thus misleading the model decision-making, the time dimension is especially considered when dividing the training set and the test set. Specifically, the division of training set and test set is based on time sequence, with the first 80% of dataset as training set and the remaining 20% as test set. Finally, 62,400 user-course interaction records are obtained for model training and 15,600 user-course interaction records for model evaluation.

In order to evaluate the contribution of different features to the course recommendation effect, Shapley Additional Explanations (SHAP) method is used to explain the model to quantify the influence of each feature on the recommendation decision. SHAP value is used to measure the contribution of each feature to the final prediction result. The greater the absolute value, the higher the importance of this feature to model decision-making. Table 2 shows the ranking of SHAP average contribution value of each feature.

Table 2 SHAP value ordering of different characteristics.
Full size table

The experimental results show that the user’s historical behavior sequence plays a central role in the course recommendation task, mainly because this feature can directly reflect the evolution trend of students’ interests and help the model accurately capture the individual needs of users. Course categories and learning preferences also have a great influence on the recommendation results, indicating that students usually refer to their own areas of interest and study habits when choosing courses. In addition, learning duration, as an important indicator to measure users’ learning engagement, also contributes greatly to the recommendation effect. However, the course score and the SHAP value of the instructor are low, which shows that students are more inclined to make decisions based on their own interests and study habits when choosing online music courses, rather than relying on external evaluation of courses or teacher information. Therefore, when optimizing the recommendation model, people should focus on user behavior characteristics and personalized interest modeling to improve the accuracy and adaptability of the recommendation system.

In the task of personalized recommendation, different recommendation methods have significant differences in feature utilization, user interest modeling and computational efficiency. In order to verify the validity of CRM-SLIE model, this study selects five representative recommendation methods for comparison, including neural collaborative filtering (NCF)27 and Deep & Cross Network (Deep & Cross)28, GRU for Recommendation (GRU4Rec)29, Sequentially Transferred Attention Model for Personalized Recommendation (STAMP)30 and Lightweight Self-Attention Networks (LightSANs)31. The characteristics of each model are as follows: NCF is a collaborative filtering method based on deep learning, which uses neural network to automatically learn the implicit representation of users and items. But it has limitations in dealing with long sequence behaviors. Deep&Cross adopts the method of combining explicit cross features with depth features, which improves the ability of feature interaction, but it is weak in dynamic sequence modeling. GRU4Rec uses GRU to model user behavior sequence, which has good effect in processing time series data, but the calculation cost is large. STAMP’s sequence recommendation method based on attention mechanism can dynamically adjust users’ interests, but its modeling ability for long sequences is limited. LightSANs’ self-attention recommendation model based on Transformer structure can efficiently model the long-term interest evolution, but the computational complexity is high. In contrast, CRM-SLIE model combines interest evolution modeling and attention mechanism, which can effectively capture the long-term interest changes of users and strike a good balance between computational efficiency and recommendation quality. Taking Area Under the Curve (AUC) and Recall@k as evaluation indexes, where @k represents the recommendation of the top k courses. The equation is as follows:

$$\:AUC=\frac{\sum\:_{j\in\:positive}{rank}_{j}-P\times\:\frac{P+1}{2}}{P\times\:N}$$
(21)
$$\:Recall@k=\frac{TP}{TP+FN}$$
(22)

\(\:{rank}_{j}\) is the ranking of positive sample j. P is the number of positive samples, that is, the number of courses clicked by users. N is the number of negative samples, that is, the number of courses that the user did not click. TP is a real example, that is, the number of courses correctly recommended by the model and clicked by users. FN is false negative, that is, the model is not recommended or recommended incorrectly, but it is actually the number of courses that users are interested in.

Experimental environment and parameters setting

The experimental environment and parameter settings are shown in Table 3.

Table 3 Experimental environment and parameter settings.
Full size table

After experimental verification, the key hyperparameters are optimized and adjusted to improve the performance and stability of the model. Specifically, the learning rate is set to 0.001, which is within the common setting range of deep learning recommendation system (0.0001 ~ 0.01). By comparing the effects of different learning rates (0.01, 0.005, 0.001, 0.0005) through experiments, it is finally determined that 0.001 can achieve the best balance between convergence speed and stability. The number of training rounds is set to 10. After experimental observation, after 10 rounds, the performance of the model tends to be stable. Continuing to increase the number of training rounds has limited improvement on AUC and may lead to over-fitting. The batch size is set to 256, and four batch sizes of 64, 128, 256 and 512 are tried in the experiment. The results show that 256 can achieve a good balance between training efficiency and model performance. The optimizer adopts Adam, because it can effectively adapt the learning rate and perform well in non-convex optimization problems, especially suitable for recommending the deep learning model of the system. The loss function selects binary cross entropy, which is mainly used for binary recommendation tasks to measure the matching degree between recommended courses and users’ actual behavior. Through the above-mentioned parameter optimization, the balanced performance of CRM-SLIE model in convergence speed, generalization ability and recommendation effect are ensured, which guarantees the reliability of subsequent experimental results and the applicability of the model.

Performance evaluation

The influence of embedding dimension on model performance

The length of student behavior sequence is set to 15, and the AUC values of different models are compared under the embedding dimensions of 32, 64 and 128, and the results are shown in Fig. 3.

Fig. 3
figure 3

AUC results of different models under different embedding dimensions.

Full size image

In Fig. 3, when the embedding dimension is 64, the AUC value of CRM-SLIE model is the highest, reaching 0.871, which is superior to other comparative models. When the embedding dimension is 128, the AUC value decreases slightly, which indicates that increasing the embedding dimension in a certain range can improve the performance of the model, but when the dimension is too large, information noise may be introduced, which will affect the performance of the model. On the whole, the performance of CRM-SLIE model in different dimensions is always better than other models, which shows that the model has stronger adaptability and robustness in the task of course recommendation. This may be because CRM-SLIE adopts GRU based on attention mechanism, which can capture the dynamic changes of students’ interest state more effectively and weight historical behaviors, making the optimization of embedded dimensions more accurate. When embedded in 64 dimensions, the model can better express the information of courses and users’ interests, and avoid the over-fitting problem caused by too high dimensions, thus improving the AUC value.

The influence of sequence length on model performance

The embedding dimension is selected as 64, and the length of student behavior sequence is set as 5, 8, 10, 15, 20, 25 and 30. The AUC of the model under different sequence lengths is compared, and the results are shown in Fig. 4.

Fig. 4
figure 4

AUC results of different models under different sequence lengths.

Full size image

Figure 4 shows that the AUC value of CRM-SLIE model keeps rising with the increase of the length of student behavior sequence, and tends to be stable after the sequence length is 20, reaching 0.872, which is the highest among all models. This shows that CRM-SLIE can effectively capture the evolution of students’ long-term interests, and it is insensitive to the change of sequence length and maintains high performance. The reason is that the CRM-SLIE model can automatically select the most valuable historical behavior for the current recommendation by weighting the students’ historical behavior through the AGRU module, so that the interest drift effect of long sequences can be effectively suppressed. In addition, the model gives different weights to different historical behaviors through the attention mechanism, so it can still maintain a high AUC value under the condition of long sequence, without information redundancy or gradient disappearance due to long sequence.

The influence of recommended quantity on model performance

The embedding dimension is 64 and the length of student behavior sequence is 30, and the recall rate of each model under different recommended numbers is evaluated. The Recall@k changes of different models are shown in Fig. 5.

Fig. 5
figure 5

Recall@k comparison of different models.

Full size image

In Fig. 5, with the increase of recommendation number k, the recall rate of CRM-SLIE model is gradually better than other comparison models, indicating that its performance in recommendation tasks has been significantly improved with the increase of recommendation number. Especially when the k value is large (such as Recall@10), the recall rate of CRM-SLIE keeps ahead, which shows that the model can better capture the changes of students’ interests and provide recommendations that match the actual needs of students. This may be due to the AGRU mechanism adopted by CRM-SLIE combined with attention weight adjustment, which can evaluate the importance of different historical behaviors, so that the model can maintain a high recall rate under different recommendation numbers. In addition, CRM-SLIE model combines dynamic interest evolution in recommendation, which can update students’ interest status in real time and make the recommendation results more accurate, so it is superior to other comparison models in Recall@5 and Recall@10.

Performance comparison of different models

The embedding dimension is 64 and the length of student behavior sequence is 30. Under the same other conditions, the results of AUC, Recall@5 and Recall@10 of each model are compared, as shown in Fig. 6.

Fig. 6
figure 6

Performance comparison results of different models.

Full size image

The results in Fig. 6 show that, under the same conditions, CRM-SLIE achieves an AUC value of 0.872, significantly higher than other models. It also demonstrates a notable lead in Recall@5 (0.262) and Recall@10 (0.364). Specifically, although GRU4Rec and STAMP have slightly higher AUC values than CRM-SLIE, CRM-SLIE significantly outperforms these models in Recall indicators. This indicates that CRM-SLIE can more accurately identify user interests and provide relevant courses in recommendation tasks. Additionally, compared to LightSANs, CRM-SLIE has a slightly higher AUC and a more pronounced advantage in Recall@10. This suggests that CRM-SLIE adapts better to increasing recommendation quantities. The main reason lies in CRM-SLIE’s use of a self-attention mechanism to capture the evolution of student interests. This enables the model to better understand long-term interest changes and make precise recommendations. At the same time, CRM-SLIE optimizes the processing of student behavior sequences, maintaining high performance even with longer sequences.

Ablation experiment

In order to verify the influence of PE and project crossing on the model performance, an ablation experiment is designed. The comparison is conducted through two experimental settings: with and without the PE module, and different item interaction methods. The PE module setting includes the PE module, enabling the capture of temporal evolution in student interests. The without PE module setting excludes the PE module, simulating the model’s performance without temporal information capture. The item interaction methods are divided into four scenarios: no interaction features, inner product interaction, Hadamard product interaction, and a combination of inner product and Hadamard product interaction. No interaction features mean no feature interaction is used, and only original features are used for modeling. Inner product interaction uses the inner product method to simulate the model capturing relationships between courses through inner products. Hadamard product interaction uses the Hadamard product method to simulate the model capturing relationships between courses through Hadamard products. Combination of inner product and Hadamard product interaction combines both methods to enhance the model’s ability to capture complex relationships between courses. The result is shown in Fig. 7.

Fig. 7
figure 7

Result of ablation experiment.

Full size image

Figure 7 shows that PE and project crossing mode have obvious influence on model performance. After adding the PE module, the model’s AUC value and Recall@10 improve compared to when the PE module is not included. This result shows that the PE module effectively captures the temporal evolution of student interests, enhancing the model’s ability to model long-term interest changes. Without the PE module, the model cannot fully consider the dynamic nature of student interests over time, leading to a decline in recommendation performance. Regarding the impact of item interaction methods, the model’s performance significantly drops when no interaction features are used. This indicates that relying solely on original features for recommendations cannot fully capture the complex relationships between courses. Item interaction methods effectively improve model performance. The combination of inner product and Hadamard product interaction achieves the best results. This suggests that the combination better captures complex relationships between courses, ultimately significantly enhancing the model’s recommendation effectiveness. The combined interaction method effectively integrates the advantages of inner product and Hadamard product, providing richer feature interactions and optimizing model performance. The ablation experiment results demonstrate that the PE module and item interaction methods significantly impact model performance. Adding the PE module notably improves the model’s ability to capture temporal dynamics. Different interaction methods vary in their ability to capture complex relationships between courses, with the combination of inner product and Hadamard product interaction performing the best and delivering the greatest performance improvement. These results indicate that combining the PE module with effective feature interaction methods is key to improving course recommendation accuracy.

Analysis of running time and calculation cost

In order to evaluate the computational efficiency of CRM-SLIE model and its scalability under different data scales, the training time and inference time of CRM-SLIE and other models in the same experimental environment are compared, and the results are shown in Fig. 8.

Fig. 8
figure 8

Training time and inference time of different models.

Full size image

From the experimental results in Fig. 8, the CRM-SLIE model shows a good balance in training time and inference time. Compared with GRU4Rec, STAMP and LightSANs, the training time of CRM-SLIE is shorter, only 85.3 s, which shows that its optimized interest evolution mechanism improves the computational efficiency. At the same time, the inference time of CRM-SLIE is 0.012 s/batch, which is better than GRU4Rec and LightSANs, indicating that its recommendation speed is faster. This study analyzes the reasons for the better calculation efficiency of CRM-SLIE model, and finds that the introduction of attention mechanism can effectively reduce redundant calculation when the model captures the dynamic changes of interest. At the same time, the structural optimization of CRM-SLIE model avoids the computational overhead caused by too deep network and improves the scalability of the model on large-scale data.

Model performance analysis

To further validate the effectiveness of the CRM-SLIE model, this section presents specific recommendation examples and analyzes the model’s performance in different scenarios. Several students’ behavior sequences are selected to demonstrate the model’s strengths and weaknesses in recommendation accuracy.

In a typical recommendation task, the CRM-SLIE model successfully recommends courses highly aligned with a student’s interests and learning progress. For a student performing well in mathematics, the model recommends several advanced math courses. These courses are closely related to the student’s past learning trajectory, including subjects like advanced algebra and calculus. By capturing the student’s long-term interest in mathematics and considering their learning progress, the model effectively recommends courses that matched the student’s interests. However, the CRM-SLIE model does not always provide ideal recommendations. In some cases, the model’s performance is less satisfactory. For students with diverse interests and scattered learning trajectories, the model’s recommendations may not meet expectations. For example, the model recommends several computer science courses to a student whose main interests were in humanities, such as literature and history. In this case, the model fails to effectively capture the student’s interest changes, leading to recommendations that did not align with the student’s actual needs. The root cause of these performance differences may lie in the diversity and dynamic nature of student interests. In some cases, students’ interests are too broad or change too quickly for the model to accurately capture these shifts. Although CRM-SLIE attempts to capture long-term interest evolution through the PE module, it may struggle to extract representative learning interest features when students’ learning trajectories are highly scattered, affecting recommendation accuracy. Additionally, the model’s recommendation effectiveness may be influenced by the course content itself. If the course content has low relevance to the student’s learning trajectory or potential interests, even if the model captures some interest features, the lack of highly relevant course data may result in poor recommendations. Therefore, further improving the model’s interest fusion mechanism in multi-interest domains could enhance its recommendation accuracy in complex interest change scenarios.

Discussion

In a word, the proposed CRM-SLIE course recommendation model based on the evolution of students’ interests can effectively capture the long-term changes of students’ interests and improve the accuracy of personalized recommendation. Compared with the existing research, the model in this study has stronger adaptability and robustness in dealing with the dynamic changes of students’ learning interests and complex course relations. Wang et al. (2021) proposed a recommendation method based on graph neural network to solve the shortcomings of existing models in explicit expression of project structural relationship and project timeliness. The experimental results showed that this model had obvious advantages in improving recommendation performance and accurately predicting learners’ preferences32. However, the computational cost of graph neural network is large, especially in the case of large data, and the time cost of model training and reasoning may increase significantly. By introducing attention mechanism and GRU structure, CRM-SLIE can capture the changes of students’ interests, effectively reduce the calculation cost and maintain high recommendation efficiency. In addition, CRM-SLIE has obvious advantages in modeling the long-term evolution of students’ interests, which can better adapt to the changes of students’ dynamic interests and has strong robustness. Jena et al. (2022) proposed an e-learning course recommendation system based on collaborative filtering mechanism, which used models such as K-nearest neighbour, singular value decomposition and collaborative filtering based on neural network to provide course selection suggestions according to users’ preferences. The experimental results showed that K- nearest neighbour performed best in hit rate and average reciprocal hit rate, with the lowest mean absolute error33. Although the K-nearest neighbors algorithm in this study performs well on some traditional metrics, collaborative filtering methods suffer from data sparsity issues and cannot effectively handle the dynamic changes in student interests. In contrast, CRM-SLIE excels in capturing the long-term evolution of student interests. It avoids the limitations of collaborative filtering in sparse data scenarios, providing more accurate personalized recommendations. Safarov et al. (2023) proposed a deep neural network recommendation method combining synchronous sequence and heterogeneous features to improve the recommendation accuracy of e-learning platform. The results showed that the recommendation accuracy of this method in Top-1 and Top-5 courses were 0.626 and 0.492 respectively34. This method improves recommendation accuracy by combining different features, but it may face challenges in feature engineering and multi-modal data fusion. Additionally, the model’s complexity is high, leading to longer training and inference times. In contrast, CRM-SLIE combines attention mechanisms and GRU. It efficiently captures student interest changes through a simplified network structure while reducing reliance on features. This makes the model more efficient in recommendation tasks while maintaining high recommendation accuracy. In contrast, CRM-SLIE can better capture students’ learning trajectory and interest changes by combining attention mechanism with GRU, thus achieving more accurate personalized recommendation. Additionally, the efficiency and adaptability of CRM-SLIE enable it to maintain strong performance across datasets of different sizes. This demonstrates greater robustness compared to other models.

Conclusion

Research contribution

In this study, the attention mechanism, GRU and project crossing module are combined to propose a CRM-SLIE model based on the evolution of students’ learning interests, aiming at improving the personalized course recommendation effect of online music learning platform. The validity of the model is verified by experiments, and the following conclusions are drawn:

  1. (1)

    Under different embedding dimensions and the length of students’ behavior sequence, the performance of CRM-SLIE model is always better than other models. When the embedding dimension is 64, the AUC value of CRM-SLIE model is the highest, and it tends to be stable after the sequence length is 20, reaching 0.872. This shows that CRM-SLIE can effectively capture the long-term interest evolution of students and has stronger adaptability and robustness in course recommendation tasks.

  2. (2)

    With the increase of recommendation number k, the recall rate of CRM-SLIE model is gradually better than other comparison models. This indicates that its performance in recommendation tasks has been significantly improved with the increase of recommendation number, which suggests that this model can better capture the changes of students’ interests and provide recommendations that are more in line with students’ actual needs.

  3. (3)

    The ablation experiment shows that PE and project crossing mode have obvious influence on the model performance. After adding PE, the AUC value and Recall@10 of the model are improved, which shows that PE can effectively capture time series information. For the project crossing mode, the combination of inner product and Hadamard product is the most helpful to capture the complex relationship among courses.

Future works and research limitations

The proposed course recommendation model effectively improves the personalized recommendation effect, but there are still the following limitations. First, the single data source may affect the generalization ability of the model. Second, the modeling of students’ interest evolution is simplified, and the changes of external environment are not fully considered. Future research can expand data sources, optimize models, and combine more real-time feedback data to improve the accuracy and adaptability of recommendations.